Unconventional reservoir compressibility evaluation method based on limit gradient lifting tree model
By using the extreme gradient boosting tree model for data preprocessing and rock mineral composition prediction, the problem of low accuracy in traditional methods is solved, enabling compressibility assessment applicable to different regions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-22
- Publication Date
- 2026-05-22
AI Technical Summary
Traditional methods lack precise data preprocessing procedures in assessing the compressibility of unconventional reservoirs, resulting in low accuracy of brittleness index prediction regression models and their inapplicability to different regions.
A rock mineral composition prediction model is established by adopting the limit gradient boosting tree model and through steps such as data cleaning, feature enhancement, normalization and dimensionality reduction. Multiple models are trained using the limit gradient algorithm and selected by evaluation index. The brittleness index is calculated for evaluation.
It improves the accuracy of the prediction model for assessing the compressibility of unconventional reservoirs, is applicable to different regions, and enables a refined assessment of formation compressibility.
Smart Images
Figure CN122072783A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of oil and gas development technology, specifically relating to an unconventional reservoir compressibility evaluation method based on the limit gradient lifting tree model. Background Technology
[0002] Currently, the evaluation methods for the fracturability of unconventional reservoirs are mainly based on the Rickman mechanical brittleness index and the Jarvie mineral empirical brittleness index. However, due to the strong heterogeneity of unconventional reservoirs, these two methods cannot obtain continuous data and both rely on a large number of core samples, resulting in high research costs and long research cycles. With the development of well logging technology, existing technologies have emerged that use experimental data and well logging data to establish regression models to predict the brittleness index, thereby evaluating fracturability.
[0003] Traditional methods collect data from the target layer of the study area and use all of this data to fit and build a regression model. If the model is not satisfactory, the formula is adjusted based on experience, ultimately resulting in a prediction model that can only be applied to a specific layer in a specific region and cannot be applied to other regions, thus having poor universality. At the same time, the data preprocessing process used in the traditional method to build the regression model is too general and lacks a precise data preprocessing process, which ultimately leads to low accuracy in building the fragile index prediction regression model. Summary of the Invention
[0004] The purpose of this invention is to provide an unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model, which solves the problem that traditional methods lack a precise data preprocessing process in the process of establishing regression models, resulting in low accuracy in establishing brittleness index prediction regression models.
[0005] The technical solution adopted in this invention is an unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model, comprising the following steps: Step 1: Extract target rock sample data and use the target rock sample data and conventional logging data as sample data for training the model based on a one-to-one correspondence of depth values. Step 2: Perform data preprocessing on the sample data from Step 1; Step 3: Use the limiting gradient algorithm to train multiple rock and mineral composition prediction models based on sample data, then select the best model through evaluation indicators and save it. Step 4: Based on the model selected in Step 3, predict the newly added production logging data and calculate the brittleness index. Evaluate the compressibility of the formation based on the brittleness index.
[0006] A further feature of this invention is that the specific process of step 1 is as follows: Based on the rock three-terminal component classification principle, the XRD whole-rock diffraction experimental data taken from the rock core were processed into three categories: clay mineral content, silica mineral content, and carbonate rock content, which were used as rock sample data. Then, conventional well logging data at the same depth as the rock core were selected to form a dataset, which was used as sample data for training the model.
[0007] In step 1, the conventional logging data consists of any one or more of GR, PE, SP, DT, resistivity, ZDEN, CNCF, M2R2, M2R3, M2R6, M2R9, M2RX, and lith.
[0008] The specific process of step 2 is as follows: Step 2.1: Perform data cleaning and feature enhancement on the sample data from Step 1, as follows: Step 2.1.1: Based on the sample data in Step 1, establish correlation scatter plots between the clay mineral content, silica mineral content, and carbonate rock content and conventional well logging curves. Step 2.1.2: Based on the correlation scatter plot obtained in Step 2.1.1, identify the points with abnormal logging responses, analyze the causes of the abnormal points, remove the abnormal points that do not conform to the lithology of the target rock, and complete the data cleaning. Step 2.1.3: Perform feature enhancement processing on the cleaned sample data, specifically using the following calculation formula:
[0009] in, It refers to the mud content in sandstone; This is a normalization process for gamma logging data; GR: natural gamma value; and These are the GR values for pure sandstone and pure mudstone, respectively, and GCUR is the volume factor for mudstone.
[0010]
[0011]
[0012] Wherein, PND is the difference between density and neutron normalization; It refers to the clay content in mudstone and shale; and These are the minimum and maximum values of density logging, respectively. and These are the minimum and maximum values for neutron logging, respectively. This is the current density logging value; This is the current value from neutron logging.
[0013]
[0014]
[0015]
[0016]
[0017] in, It refers to the content of siliceous minerals in mudstone and shale; It is the maximum time difference of sound waves; , It is the slope where the sound wave and density intersect; It is the slope where the neutron and density intersect; It is the slope of the intersection of the sound wave and the neutron. This is the current value of the sound wave time difference. The values are the current density logging values, and a, b, c, and d are correction coefficients.
[0018] The feature data is normalized using the following formula:
[0019] in, For the current feature data, For the normalized data, This is the minimum value in this feature data. This is the maximum value in this feature data.
[0020] Step 2.1.4: Select feature vectors and perform dimensionality reduction on the normalized feature dataset.
[0021] The specific process of step 2.1.4 is as follows: Calculate the SHAP value for each sample of each feature provided to the model, sort the features that affect the model results according to the SHAP values, and perform dimensionality reduction judgment and processing on the feature dataset based on the sorting results of the features.
[0022] The specific process of step 3 is as follows: Step 3.1: Randomly select 70% of the data preprocessed in Step 2 as the model training set, and use the remaining 30% of the data as the model test set; Step 3.2: Based on the training dataset, select several hyperparameters to maximize model performance and prevent overfitting. Use GridSearchCV to set the optimal hyperparameter set and use 10x cross-validation to test the stability of model performance. Step 3.3: Using the extreme gradient boosting machine learning algorithm, at least 20 different models are generated to evaluate the impact of randomness on model performance, thereby establishing prediction models for clay mineral content, silica mineral content, and carbonate rock content, respectively. Step 3.4: Verify the accuracy of the model using the test set. Utilize the model evaluation metrics RMSE and R2 to evaluate the performance of the clay mineral content prediction model, silica mineral content prediction model, and carbonate rock content prediction model. If the R2 value of the model is greater than 0.8 and the RMSE is less than 1, it passes the evaluation metrics and is saved as the preferred model. If the R2 value is less than 0.8 or the RMSE value is greater than 1, the model evaluation metrics are not met. Perform feature vector contribution analysis on the model, delete features that contribute less to the model, and then return to step 2 above to perform feature enhancement again before establishing the corresponding prediction model.
[0023] The specific calculation process for the model evaluation metrics RMSE and R2 is as follows: The formula for calculating the model evaluation metric RMSE is as follows:
[0024] The formula for calculating the evaluation index R2 is as follows:
[0025] in, The clay mineral content, silica mineral content, and carbonate rock content were obtained from indoor tests on a sample set. The model predicts the clay mineral content, silica mineral content, and carbonate rock content. The average value of the model predictions is represented by i, where i represents the first model and N represents the last model.
[0026] The specific process of step 4 is as follows: Step 4.1: Using the clay mineral content, silica mineral content, and carbonate rock content prediction models saved in Step 3, predict the three-terminal component composition of the newly added production logging data. Step 4.2: Based on the predicted clay mineral content, silica mineral content, and carbonate rock content, calculate the brittleness index, and then evaluate the compressibility of the formation based on the brittleness index. The formula for calculating the brittleness index is as follows:
[0027] Calculate the mineral heterogeneity index for each measurement point:
[0028] in, It is the brittle mineral index. It is a mineral heterogeneity index. It refers to the content of siliceous minerals. It refers to the carbonate content. It refers to the clay mineral content; Well logging data was recorded every 0.25m as one logging point, and the fracturing index of each logging point was calculated based on the heterogeneity index and brittleness index.
[0029] in, It is the fracturing index. The larger the value, the higher the compressibility. 'a' is the brittle mineral index coefficient, and 'b' is the mineral heterogeneity index coefficient.
[0030] The beneficial effects of this invention are This invention presents an unconventional reservoir compressibility evaluation method based on the extreme gradient boosting tree model. By designing specific data preprocessing procedures such as data cleaning, feature enhancement, normalization, and dimensionality reduction, the sample data is processed, and then a prediction model is established based on the processed sample data, which effectively improves the accuracy of the prediction model. Attached Figure Description
[0031] Figure 1 This is the flowchart of the unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model of the present invention; Figure 2 This is a scatter plot showing the correlation between siliceous mineral content and conventional logging curves in the unconventional reservoir compressibility evaluation method based on the ultimate gradient lifting tree model of this invention. Figure 3 This is the basis index diagram for the data dimensionality reduction processing diagram in the unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model of this invention; Figure 4 This is a comparison chart of predicted and actual values of unconventional reservoir rock and minerals in the unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model of this invention; Figure 5 This is a graph showing the predicted results of clay mineral content, silica mineral content, and carbonate rock content in the unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model of this invention. Detailed Implementation
[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.
[0033] This invention relates to an unconventional reservoir compressibility evaluation method based on the extreme gradient boosting tree model, such as... Figure 1 As shown, it includes the following steps: Step 1: Extract target rock sample data and use the target rock sample data and conventional logging data as sample data for training the model based on a one-to-one correspondence of depth values. The specific process of step 1 is as follows: Based on the rock three-terminal component classification principle, the XRD whole-rock diffraction experimental data taken from the rock core were processed into three categories: clay mineral content, silica mineral content, and carbonate rock content, which were used as rock sample data. Then, conventional well logging data at the same depth as the rock core were selected to form a dataset, which was used as sample data for training the model.
[0034] The conventional logging data in step 1 includes: Natural gamma (GR) logging data, photoelectric section index (PE) logging data, spontaneous potential (SP) logging data, sonic transit time (DT) logging data, resistivity logging data, density (ZDEN) logging data, density slope (TN), neutron (CNCF) logging data, 5 array induced resistivity logging data (M2R2, M2R3, M2R6, M2R9, M2RX), and lithological stratification data (lith).
[0035] Step 2: Perform data preprocessing on the sample data from Step 1; The specific process of step 2 is as follows: Step 2.1: Perform data cleaning and feature enhancement on the sample data from Step 1, as follows: Step 2.1.1: Based on the sample data from Step 1, establish correlation scatter plots between the clay mineral content, silica mineral content, and carbonate rock content and conventional logging curves. For example, using 6 conventional logging curves will produce a total of 18 scatter plots.
[0036] Step 2.1.2: Based on the correlation scatter plot obtained in Step 2.1.1, identify the points with abnormal logging responses, analyze the causes of the abnormal points, remove the abnormal points that do not conform to the lithology of the target rock, and complete the data cleaning. For example: Figure 2 As shown, Figure 2 The scatter plots showing the correlation between CNCF logging and ZDEN logging and quartz content are presented. Figure 2 In the figure, 'a' indicates that the correlation coefficient between the measured quartz content and CNCF logging before data cleaning was 0.41, and... Figure 2 In the text "b", it indicates that the density logging data range before data cleaning was 1.3–2.8. After data cleaning, as shown below... Figure 2 c and Figure 2 As shown in d, c indicates that the correlation between the measured quartz content and CNCF logging data has increased to 0.53, and d indicates that the distribution range of density logging data is focused at 2.2-2.8.
[0037] Step 2.1.3: Perform feature enhancement processing on the cleaned sample data. This involves using well logging data such as gamma ray, neutron, density, and sonic transit time to calculate the mud content, clay content in shale, and silica mineral content in shale, which are then used as new feature vectors to achieve feature enhancement. The specific calculation formula is as follows: (1) (2) in, It refers to the mud content in sandstone; This is a normalization process for gamma logging data; GR: natural gamma value; and These are the GR values for pure sandstone and pure mudstone, respectively, and GCUR is the volume factor for mudstone.
[0038] (3) (4) Wherein, PND is the difference between density and neutron normalization; It refers to the clay content in mudstone and shale; and These are the minimum and maximum values for density logging, respectively. and These are the minimum and maximum values for neutron logging, respectively. This is the current density logging value; This is the current value from neutron logging.
[0039] (5) (6) (7) (8) in, It refers to the content of siliceous minerals in mudstone and shale; It is the maximum time difference of sound waves; , It is the slope where the sound wave and density intersect; It is the slope where the neutron and density intersect; It is the slope of the intersection of the sound wave and the neutron. This is the current value of the sound wave time difference. The values are the current density logging values, and a, b, c, and d are correction coefficients.
[0040] The feature data is normalized using the following formula: (9) in, For the current feature data, For the normalized data, This is the minimum value in this feature data. This is the maximum value in this feature data.
[0041] Step 2.1.4: Perform feature vector selection and dimensionality reduction on the normalized feature dataset, specifically as follows: For each sample of each feature provided to the model, the SHAP value (SHapley Additive ex Planations) is calculated, and the features that affect the model results are sorted according to the SHAP value, that is, the SHAP value is sorted from high to low. The feature dataset is then subjected to dimensionality reduction judgment and processing based on the sorting results of the features.
[0042] Here, the SHAP value represents the construction of a set of participants containing all feature subsets, the calculation of the marginal contribution of each feature subset, the calculation of features based on the marginal contribution, and the higher the SHAP value, the greater the contribution of the feature to the model. By observing the SHAP distribution characteristics, the inflection point where the SHAP value drops significantly is marked, and the feature vector after the inflection point is deleted, thereby completing the dimensionality reduction.
[0043] By designing specific data preprocessing procedures such as data cleaning, feature enhancement, normalization, and dimensionality reduction, the sample data is processed. Then, a prediction model is built based on the processed sample data, effectively improving the accuracy of the prediction model. For example, the distribution of SHAP values of feature vectors in different datasets... Figure 3 As shown. Figure 3 In the figure, 'a' represents the original data, which contains 12 feature vectors. At this point, the model accuracy is 0.46. CNCF logging and PE logging contribute significantly to the model, followed by GR logging and ZDEN logging. Figure 3 As shown in Figure b, after data cleaning, the model accuracy can reach 0.6. CNCF logging significantly enhances its contribution to the model, followed by GR logging. Figure 3 In the figure, c represents the 19 feature vectors obtained after feature enhancement, and the model accuracy is 0.81. At this point, CNCF and TN logging have the greatest contribution to the model, indicating that CNCF and TN logging are the main factors affecting the prediction model of siliceous minerals or clay minerals.
[0044] Step 3: Use the limiting gradient algorithm to train multiple rock and mineral composition prediction models based on sample data, then select the best model through evaluation indicators and save it. The specific process of step 3 is as follows: Step 3.1: Randomly select 70% of the data preprocessed in Step 2 as the model training set, and use the remaining 30% of the data as the model test set; The data after preprocessing in step 2 is divided into two parts: 70% is used to build the model and 30% is used to validate the model. When the accuracy of the validation model is close to 1, it means that the regression model has high accuracy and is usable. Only when a validation set is preset when the model is built and the accuracy of the validation set is high can it be said that the model has universality and is applicable not only to this region but also to other regions.
[0045] Step 3.2: Based on the training dataset, select several hyperparameters to maximize model performance and prevent overfitting. Use GridSearchCV to set the optimal hyperparameter set and use 10x cross-validation to test the stability of model performance. Step 3.3: Using the extreme gradient boosting machine learning algorithm, at least 20 different models are generated to evaluate the impact of randomness on model performance, thereby establishing clay mineral content prediction models, silica mineral content prediction models, and carbonate rock content prediction models respectively. In evaluating the impact of randomness on model performance by generating at least 20 different models, the differences in training datasets, initial weights, and hyperparameter settings lead to varying results from each machine learning algorithm, which is considered randomness. If, assuming 30 models are set up, the error values of these 30 models are very small, it indicates that the model is relatively stable, meaning the randomness of the machine learning algorithm itself has a small impact on the model's performance, and the model is relatively stable and usable.
[0046] Step 3.4: Verify the accuracy of the model using the test set. Evaluate the performance of the clay mineral content prediction model, silica mineral content prediction model, and carbonate rock content prediction model using the model evaluation indicators root mean square error (RMSE) and correlation coefficient (R2). If the R2 value of the model is greater than 0.8 and the RMSE is less than 1, it passes the evaluation indicators and is saved as the preferred model. If the R2 value is less than 0.8 or the RMSE value is greater than 1, it does not meet the model evaluation indicators. Perform feature vector contribution analysis on the model, delete features that contribute less to the model, and then return to step 2 above to perform feature enhancement again before establishing the corresponding prediction model.
[0047] The formula for calculating the model evaluation metric RMSE is as follows: (10) The formula for calculating the evaluation index R2 is as follows: (11) in, The clay mineral content, silica mineral content, and carbonate rock content were obtained from indoor tests on a sample set. The model predicts the clay mineral content, silica mineral content, and carbonate rock content. The average value of the model predictions is represented by i, where i represents the first model and N represents the last model.
[0048] like Figure 4 As shown, taking siliceous mineral content as an example, this is a scatter plot of the prediction model results based on different datasets. Each dataset was iterated 30 times. After data cleaning and feature enhancement, Figure 4 In this context, 'a' indicates that the RMSE value gradually decreases and its distribution range gradually shrinks. Figure 4 In the diagram, 'b' indicates that the R² value gradually increases and its distribution range gradually decreases, suggesting that the model's accuracy and stability are gradually improving. For example... Figure 5 As shown, when the silica mineral content prediction model was applied to 12 wells in the work area, the correlation coefficient between the measured data and the calculated prediction data reached 0.93.
[0049] Step 4: Based on the model selected in Step 3, predict the newly added production logging data and calculate the brittleness index. Evaluate the compressibility of the formation based on the brittleness index.
[0050] The specific process of step 4 is as follows: Step 4.1: Using the clay mineral content, silica mineral content, and carbonate rock content prediction models saved in Step 3, predict the three-terminal component composition of the newly added production logging data. Step 4.2: Based on the predicted clay mineral content, silica mineral content, and carbonate rock content, calculate the brittleness index, and then evaluate the compressibility of the formation based on the brittleness index. The formula for calculating the brittleness index is as follows: (12) Calculate the mineral heterogeneity index for each measurement point: (13) in, It is the brittle mineral index. It is a mineral heterogeneity index. It refers to the content of siliceous minerals. It refers to the carbonate content. It refers to the clay mineral content.
[0051] Log data was recorded every 0.25m as one logging point, and the fracturing index of each logging point was calculated based on the heterogeneity index and brittleness index. (14) in, It is the fracturing index. The larger the value, the higher the compressibility; a is the brittle mineral index coefficient, and b is the mineral heterogeneity index coefficient.
[0052] This invention relates to an unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model, and specific embodiments are as follows: Example 1 An unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model includes the following steps: Step 1: Extract target rock sample data and use the target rock sample data and conventional logging data as sample data for training the model based on a one-to-one correspondence of depth values. Step 2: Perform data preprocessing on the sample data from Step 1; Step 3: Use the limiting gradient algorithm to train multiple rock and mineral composition prediction models based on sample data, then select the best model through evaluation indicators and save it. Step 4: Based on the model selected in Step 3, predict the newly added production logging data and calculate the brittleness index. Evaluate the compressibility of the formation based on the brittleness index.
[0053] Example 2 An unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model includes the following steps: Step 1: Extract target rock sample data and use the target rock sample data and conventional logging data as sample data for training the model based on a one-to-one correspondence of depth values. The specific process of step 1 is as follows: Based on the rock three-terminal component classification principle, the XRD whole-rock diffraction experimental data taken from the rock core were processed into three categories: clay mineral content, silica mineral content, and carbonate rock content, which were used as rock sample data. Then, conventional well logging data at the same depth as the rock core were selected to form a dataset, which was used as sample data for training the model.
[0054] In step 1, the conventional logging data consists of any one or more of GR, PE, SP, DT, resistivity, ZDEN, CNCF, M2R2, M2R3, M2R6, M2R9, M2RX, and lith.
[0055] Step 2: Perform data preprocessing on the sample data from Step 1; Step 2.1: Perform data cleaning and feature enhancement on the sample data from Step 1, as follows: Step 2.1.1: Based on the sample data in Step 1, establish correlation scatter plots between the clay mineral content, silica mineral content, and carbonate rock content and conventional well logging curves. Step 2.1.2: Based on the correlation scatter plot obtained in Step 2.1.1, identify the points with abnormal logging responses, analyze the causes of the abnormal points, remove the abnormal points that do not conform to the lithology of the target rock, and complete the data cleaning. Step 2.1.3: Perform feature enhancement processing on the cleaned sample data, specifically using the following calculation formula:
[0056]
[0057] in, It refers to the mud content in sandstone; This is a normalization process for gamma logging data; GR: natural gamma value; and These are the GR values for pure sandstone and pure mudstone, respectively, and GCUR is the volume factor for mudstone.
[0058]
[0059]
[0060] Wherein, PND is the difference between density and neutron normalization; It refers to the clay content in mudstone and shale; and These are the minimum and maximum values for density logging, respectively. and These are the minimum and maximum values for neutron logging, respectively. This is the current density logging value; This is the current value from neutron logging.
[0061]
[0062]
[0063]
[0064]
[0065] in, It refers to the content of siliceous minerals in mudstone and shale; It is the maximum time difference of sound waves; , It is the slope where the sound wave and density intersect; It is the slope where the neutron and density intersect; It is the slope of the intersection of the sound wave and the neutron. This is the current value of the sound wave time difference. The values are the current density logging values, and a, b, c, and d are correction coefficients.
[0066] The feature data is normalized using the following formula:
[0067] in, For the current feature data, For the normalized data, This is the minimum value in this feature data. This is the maximum value in this feature data.
[0068] Step 2.1.4: Perform feature vector selection and dimensionality reduction on the normalized feature dataset, specifically as follows: Calculate the SHAP value for each sample of each feature provided to the model, sort the features that affect the model results according to the SHAP values, and perform dimensionality reduction judgment and processing on the feature dataset based on the sorting results of the features.
[0069] Step 3: Use the limiting gradient algorithm to train multiple rock and mineral composition prediction models based on sample data, then select the best model through evaluation indicators and save it. Step 3.1: Randomly select 70% of the data preprocessed in Step 2 as the model training set, and use the remaining 30% of the data as the model test set; Step 3.2: Based on the training dataset, select several hyperparameters to maximize model performance and prevent overfitting. Use GridSearchCV to set the optimal hyperparameter set and use 10x cross-validation to test the stability of model performance. Step 3.3: Using the extreme gradient boosting machine learning algorithm, at least 20 different models are generated to evaluate the impact of randomness on model performance, thereby establishing prediction models for clay mineral content, silica mineral content, and carbonate rock content, respectively. Step 3.4: Verify the accuracy of the model using the test set. Utilize the model evaluation metrics, Root Mean Square Error (RMSE) and Correlation Coefficient (R²), to evaluate the performance of the clay mineral content prediction model, silica mineral content prediction model, and carbonate rock content prediction model. If the R² value of the model is greater than 0.8 and the RMSE is less than 1, it passes the evaluation metrics and is saved as the preferred model. If the R² value is less than 0.8 or the RMSE value is greater than 1, the model evaluation metrics are not met. Perform eigenvector contribution analysis on the model, delete features that contribute less to the model, and then return to Step 2 above to perform feature enhancement again before establishing the corresponding prediction model. The formula for calculating the model evaluation metric RMSE is as follows:
[0070] The formula for calculating the evaluation index R2 is as follows:
[0071] in, The clay mineral content, silica mineral content, and carbonate rock content were obtained from indoor tests on a sample set. The model predicts the clay mineral content, silica mineral content, and carbonate rock content. The average value of the model predictions is represented by i, where i represents the first model and N represents the last model.
[0072] Step 4: Based on the model selected in Step 3, predict the newly added production logging data and calculate the brittleness index. Evaluate the compressibility of the formation based on the brittleness index.
[0073] Step 4.1: Using the clay mineral content, silica mineral content, and carbonate rock content prediction models saved in Step 3, predict the three-terminal component composition of the newly added production logging data. Step 4.2: Based on the predicted clay mineral content, silica mineral content, and carbonate rock content, calculate the brittleness index, and then evaluate the compressibility of the formation based on the brittleness index. The formula for calculating the brittleness index is as follows:
[0074] Calculate the mineral heterogeneity index for each measurement point:
[0075] in, It is the brittle mineral index. It is a mineral heterogeneity index. It refers to the content of siliceous minerals. It refers to the carbonate content. It refers to the clay mineral content; Log data was recorded every 0.25m as one logging point, and the fracturing index of each logging point was calculated based on the heterogeneity index and brittleness index.
[0076] in, It is the fracturing index. The larger the value, the higher the compressibility. 'a' is the brittle mineral index coefficient, and 'b' is the mineral heterogeneity index coefficient.
[0077] Example 3 Twelve exploration wells were selected from a target area of a certain oil reservoir. 184 shale core samples were taken from different depths of nine wells for XRD whole-rock diffraction experiments, determining the contents of siliceous minerals, clay minerals, and carbonate minerals. Simultaneously, the actual well depth corresponding to each core sample was marked, obtaining 12 conventional logging curves for that depth, including GR, PE, SP, DT, ZDEN, CNCF, five array induced resistivity lines (M2R2, M2R3, M2R6, M2R9, M2RX), and lith. Using siliceous mineral content, clay mineral content, and carbonate mineral content as target values, and the 12 conventional logging data as feature values, a 184×13 matrix was established as the sample dataset. Due to the strong heterogeneity of shale, the data was cleaned according to the steps in Example 2 above, and feature enhancement was performed on the cleaned data, resulting in a 184×19 matrix. Of the 184 datasets, 128 were used as the training set to build prediction models for the content of siliceous minerals, clay minerals, and carbonate minerals, respectively. The remaining 56 datasets were used as the test set to validate the models. The results show that the R-values of the models are [data missing]. 2 The model's performance index was 0.81, and its RMSE was 0.9, meeting the model evaluation criteria. The model was applied to 12 wells in the study area, obtaining continuous distributions of siliceous minerals, clay minerals, and carbonate minerals at 0.25m intervals in the target layer of each well. Based on this, the fracturing index was calculated, providing a valuable reference for the detailed evaluation of the vertical compressibility of single wells.
[0078] Example 4 The selected reservoir is a low-permeability to ultra-low-permeability reservoir. After long-term injection-production development, the declining production rate has been increasing year by year. To further improve production, in-situ perforation fracturing is necessary. Due to years of development in the old area, experimental analysis and testing data are lacking, and fracturing capability evaluation has not been conducted. The well and layer selection scheme is unclear, hindering the implementation of specific measures. Therefore, experimental data of 2650 rock and mineral components from over 300 wells in the entire area were collected and matched one-to-one with conventional logging data at the corresponding depths. Conventional logging data included GR, SP, DT, ZDEN, and CNCF, serving as the sample dataset. Since the target layer is a tight sandstone reservoir with a relatively stable sedimentary environment, logging technology exhibits significant response characteristics to this type of formation. Therefore, no data preprocessing or data augmentation procedures were performed. Following the steps in Example 2 above, a predictive model for the content of siliceous minerals, clay minerals, and carbonate minerals can be directly established. Similarly, 1855 sets of data were used as the training set to build the model, and 795 sets of data were used as the test set to validate the model. The evaluation index R for validating the model is... 2 The value was 0.86, and the RMSE was 0.5. The model was applied to more than 300 wells in the study area to obtain the continuous distribution of siliceous minerals, clay minerals and carbonate rock minerals at a target layer interval of 0.25m in each well. Based on this, the fracturing index was calculated, which provides a useful reference for the fine evaluation of the vertical compressibility of single wells.
[0079] Example 5 Shale oil fields were selected because the lithology of these reservoirs is complex, mainly consisting of fine-grained sediments such as shale, mudstone, and siltstone. Conventional logging curves show indistinct response characteristics, hindering well and layer selection for fracturing operations. Therefore, XRD experimental data and corresponding conventional logging data for different lithologies were collected as a sample dataset. Following the steps in Example 2 above, data cleaning and feature enhancement were performed to establish a predictive model for the content of siliceous minerals, clay minerals, and carbonate minerals. This model was then used to calculate the fracturing capability of small layers, providing a reference for well and layer selection for fracturing measures. The model evaluation index R... 2 =0.78, RMSE=1.1, which basically meets the requirements of the prediction model, indicating that the model establishment method is also applicable to the evaluation of the compressibility of complex lithological strata.
[0080] Example 6 The database collected XRD whole-rock diffraction data from 98 core samples from 8 exploration wells, along with GR, PE, SP, DT, ZDEN, CNCF, 5 array induced resistivity (M2R2, M2R3, M2R6, M2R9, M2RX) data from the same depth, and 12 conventional logging curves for lithology. Using the silica, clay, and carbonate mineral contents from the XRD data as target values and the 12 conventional logging data as feature values, a 98×13 matrix was constructed as the sample dataset. The cleaned data was processed according to the steps in Example 2 above, and feature enhancement was performed, resulting in a 98×19 matrix. 71 of the 98 datasets were used as the training set to build prediction models for silica, clay, and carbonate mineral contents, respectively. The remaining 27 datasets were used as the test set to validate the models. The results show that the model's R... 2 The model's performance index was 0.82, and its RMSE was 0.8, meeting the model evaluation criteria. The model was then applied to 10 other wells in the study area where XRD experimental data were not collected. This yielded a continuous distribution of siliceous minerals, clay minerals, and carbonate minerals at 0.25m intervals in the target layer of each well. Based on this, the fracturing index was calculated, providing a valuable reference for the detailed evaluation of the vertical compressibility of single wells.
[0081] As can be seen from the above examples 3-6, the unconventional reservoir compressibility evaluation method based on the limit gradient lifting tree model of the present invention is applicable to different regions such as low-permeability to ultra-low-permeability reservoirs and shale oil fields. By calculating the fracturing index, it provides a useful reference for the fine evaluation of the vertical compressibility of a single well and has strong universality.
Claims
1. A method for evaluating the compressibility of unconventional reservoirs based on a limit gradient boosting tree model, characterized in that, Includes the following steps: Step 1: Extract target rock sample data and use the target rock sample data and conventional logging data as sample data for training the model based on a one-to-one correspondence of depth values. Step 2: Perform data preprocessing on the sample data from Step 1; Step 3: Use the limiting gradient algorithm to train multiple rock and mineral composition prediction models based on sample data, then select the best model through evaluation indicators and save it. Step 4: Based on the model selected in Step 3, predict the newly added production logging data and calculate the brittleness index. Evaluate the compressibility of the formation based on the brittleness index.
2. The unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model according to claim 1, characterized in that, The specific process of step 1 is as follows: Based on the rock three-terminal component classification principle, the XRD whole-rock diffraction experimental data taken from the rock core were processed into three categories: clay mineral content, silica mineral content, and carbonate rock content, which were used as rock sample data. Then, conventional well logging data at the same depth as the rock core were selected to form a dataset, which was used as sample data for training the model.
3. The unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model according to claim 1, characterized in that, The conventional logging data mentioned in step 1 consists of any one or more of GR, PE, SP, DT, resistivity, ZDEN, CNCF, M2R2, M2R3, M2R6, M2R9, M2RX, and lith.
4. The unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model according to claim 1, characterized in that, The specific process of step 2 is as follows: Step 2.1: Perform data cleaning and feature enhancement on the sample data from Step 1, as follows: Step 2.1.1: Based on the sample data in Step 1, establish correlation scatter plots between the clay mineral content, silica mineral content, and carbonate rock content and conventional well logging curves. Step 2.1.2: Based on the correlation scatter plot obtained in Step 2.1.1, identify the points with abnormal logging responses, analyze the causes of the abnormal points, remove the abnormal points that do not conform to the lithology of the target rock, and complete the data cleaning. Step 2.1.3: Perform feature enhancement processing on the cleaned sample data, specifically using the following calculation formula: (1) (2) in, It refers to the mud content in sandstone; This is a normalization process for gamma logging data; GR: natural gamma value; and These are the GR values for pure sandstone and pure mudstone, respectively, and GCUR is the volume factor for mudstone. (3) (4) Wherein, PND is the difference between density and neutron normalization; It refers to the clay content in mudstone and shale; and These are the minimum and maximum values for density logging, respectively. and These are the minimum and maximum values for neutron logging, respectively. This is the current density logging value; This is the current value from neutron logging; (5) (6) (7) (8) in, It refers to the content of siliceous minerals in mudstone and shale; It is the maximum time difference of sound waves; , It is the slope where the sound wave and density intersect; It is the slope where the neutron and density intersect; It is the slope of the intersection of the sound wave and the neutron. This is the current value of the sound wave time difference. The values are the current density logging values, and a, b, c, and d are correction coefficients. The feature data is normalized using the following formula: (9) in, For the current feature data, For the normalized data, This is the minimum value in this feature data. This is the maximum value in this feature data; Step 2.1.4: Select feature vectors and perform dimensionality reduction on the normalized feature dataset.
5. The unconventional reservoir compressibility evaluation method based on the limit gradient boosting tree model according to claim 4, characterized in that, The specific process of step 2.1.4 is as follows: Calculate the SHAP value for each sample of each feature provided to the model, sort the features that affect the model results according to the SHAP values, and perform dimensionality reduction judgment and processing on the feature dataset based on the sorting results of the features.
6. The method for evaluating the compressibility of unconventional reservoirs based on the limiting gradient boosting tree model according to claim 1, characterized in that, The specific process of step 3 is as follows: Step 3.1: Randomly select 70% of the data preprocessed in Step 2 as the model training set, and use the remaining 30% of the data as the model test set; Step 3.2: Based on the training dataset, select several hyperparameters to maximize model performance and prevent overfitting. Use GridSearchCV to set the optimal hyperparameter set and use 10x cross-validation to test the stability of model performance. Step 3.3: Using the extreme gradient boosting machine learning algorithm, at least 20 different models are generated to evaluate the impact of randomness on model performance, thereby establishing prediction models for clay mineral content, silica mineral content, and carbonate rock content, respectively. Step 3.4: Verify the accuracy of the model using the test set. Utilize the model evaluation metrics RMSE and R2 to evaluate the performance of the clay mineral content prediction model, silica mineral content prediction model, and carbonate rock content prediction model. If the R2 value of the model is greater than 0.8 and the RMSE is less than 1, it passes the evaluation metrics and is saved as the preferred model. If the R2 value is less than 0.8 or the RMSE value is greater than 1, the model evaluation metrics are not met. Perform feature vector contribution analysis on the model, delete features that contribute less to the model, and then return to step 2 to perform feature enhancement again before establishing the corresponding prediction model.
7. The method for evaluating the compressibility of unconventional reservoirs based on the ultimate gradient boosting tree model according to claim 6, characterized in that, The specific calculation process for the model evaluation metrics RMSE and R2 is as follows: The formula for calculating the model evaluation metric RMSE is as follows: (10) The formula for calculating the evaluation index R2 is as follows: (11) in, The clay mineral content, silica mineral content, and carbonate rock content were obtained from indoor tests on a sample set. The model predicts the clay mineral content, silica mineral content, and carbonate rock content. The average value of the model predictions is represented by i, where i represents the first model and N represents the last model.
8. The method for evaluating the compressibility of unconventional reservoirs based on the limiting gradient boosting tree model according to claim 1, characterized in that, The specific process of step 4 is as follows: Step 4.1: Using the clay mineral content, silica mineral content, and carbonate rock content prediction models saved in Step 3, predict the three-terminal component composition of the newly added production logging data. Step 4.2: Based on the predicted clay mineral content, silica mineral content, and carbonate rock content, calculate the brittleness index, and then evaluate the compressibility of the formation based on the brittleness index. The formula for calculating the brittleness index is as follows: (12) Calculate the mineral heterogeneity index for each measurement point: (13) in, It is the brittle mineral index. It is a mineral heterogeneity index. It refers to the content of siliceous minerals. It refers to the carbonate content. It refers to the clay mineral content; Log data was recorded every 0.25m as one logging point, and the fracturing index of each logging point was calculated based on the heterogeneity index and brittleness index. (14) in, It is the fracturing index. The larger the value, the higher the compressibility. 'a' is the brittle mineral index coefficient, and 'b' is the mineral heterogeneity index coefficient.